# An interpretable machine learning algorithm enables dynamic 48-hour mortality prediction during an ICU stay

## TL;DR

A machine learning model was developed to dynamically predict ICU patient mortality within 48 hours, updating daily and showing strong performance across datasets.

## Contribution

The model introduces dynamic, interpretable 48-hour mortality prediction using time-varying SHAP values for ICU patients.

## Key findings

- The model achieved AUROCs of 0.909 in training and 0.886 in testing datasets.
- External validation on MIMIC-IV showed an AUROC of 0.859.
- Time-varying SHAP values enhance interpretability and reflect changes in patient status.

## Abstract

Accurate short-term mortality prediction is essential for optimizing ICU management and improving patient outcomes. Many existing models rely on static data and do not reflect the dynamic progression of critical illness. This study aimed to develop and validate an interpretable machine learning algorithm that enables dynamic 48-hour mortality prediction throughout the ICU stay.

We conducted a retrospective cohort study using electronic health records of 9,786 ICU patients treated between 2018 and 2022 at a German university hospital. A machine learning model was developed to predict 48-hour mortality, updated every 24 hours during the ICU stay. We trained and evaluated a Light Gradient-Boosting Machine using nested cross-validation and assessed performance via area under the receiver operating characteristic curve. External validation was performed on the MIMIC-IV database. Feature importance was analyzed using SHAP values.

Here, we show that the Light Gradient-Boosting Machine algorithm (LGBM-48h) achieves AUROCs of 0.909 (95% CI: 0.901–0.917) in the training and 0.886 (95% CI: 0.878–0.895) in the testing dataset. External validation using the MIMIC-IV database yields an AUROC of 0.859 (95% CI: 0.849–0.870). The model enables effective risk stratification across the ICU stay and reflects individual changes in patient status over time. Time-varying SHAP values improve interpretability by highlighting associated features.

LGBM-48h provides a dynamic and interpretable framework for short-term ICU mortality prediction. The model may support clinical decision-making and prioritization of care, but requires further validation in real-time and prospective settings.

In this study, we developed a computer algorithm to help doctors predict whether a patient in the intensive care unit (ICU) may die within the next 48 hours. This can support timely treatment decisions and improve patient care. We used medical records from nearly 10,000 ICU patients to train the algorithm and tested it on data from another hospital. The algorithm updates its prediction every day using new patient information. We found that the tool was able to reliably identify high-risk patients and adapt to changes in their condition. It also shows which clinical values most influence the risk. This kind of technology could help ICU teams better plan treatments and use resources, especially in very busy hospital environments.

Britsch et al. develop a dynamic machine learning model that predicts 48-hour mortality for intensive care unit patients throughout their stay. The model shows robust performance across disease groups and timepoints, offering interpretable, clinically relevant risk categorization.